Prediction accuracy

How we measure accuracy

Every forecast is frozen in our prediction log before the match kicks off. After the final whistle, we compare the frozen prediction to the real result and record:

• Outcome accuracy — did we call the right winner (or draw)?
• Goal-margin accuracy — did we predict the exact goal differential?
• Margin error (MAE) — how many goals off was our margin estimate on average?
• Calibration (Brier score, RPS, log loss) — were our confidence levels well-calibrated or overconfident?

Nothing is staged or cherry-picked. Every graded prediction is published in the predictions log.

Model comparison

We back-test our model head-to-head against every serious alternative — Dixon-Coles, a self-fitting regression (GLM), generic Elo, raw goal-scoring strength, and simply sorting by national rank. Each approach gets scored on the same real decisive games. Our model has come out on top in the states where we've validated it.